How to generate un-watermarked text while still using AI

How to generate un-watermarked text while still using AI

The reel hands you one: led → managed, the anchor's payoff. Two words, warm ink on cream, one terracotta accent on the swapped word. No face, no arrows, no shouting. It reads at phone size and it is literally the argument.

Second option: "no copy button" in serif, with a crossed-out copy icon.

Chapters — verified against measured beat durations 0:00 The question 0:11 "Led a team of six" 0:27 Not a taste problem 0:48 There's no copy button 1:06 Where the watermark comes in 1:29 What it proves, and what it doesn't 1:59 Your turn

Every segment clears YouTube's 10-second minimum. First chapter starts at 0:00 as required.

Tags AI editing, LLM editing, AI rewrote my text, watermarking, text watermark, AI writing tools, editing fidelity, autoregressive models, résumé AI, AI proofreading, prompt engineering, Claude, computational skepticism

What this video covers

• 0:00 — The question
• 0:11 — "Led a team of six"
• 0:27 — Not a taste problem
• 0:48 — There's no copy button
• 1:06 — Where the watermark comes in
• 1:29 — What it proves, and what it doesn't
• 1:59 — Your turn

Transcript

A lightly formatted transcript of the video, with timestamps so you can jump to any section on YouTube.

The question

[0:00] Someone asked why an AI editor changes words it was never asked to touch. The answer isn't what you'd guess. Liam, take them through it.

"Led a team of six"

[0:11] >> Ask an AI to fix one typo on your resume and it can quietly change a word you chose on purpose. Not because the new word is worse, often it's fine, but because nothing in the system was stopping it. Hold on to one line, let a team of six. You picked lead, you meant led.

Not a taste problem

[0:27] The obvious explanation is that the model judged your word and preferred its own, a taste problem fixable by telling it to change nothing else. But it does this even when told not to, and the part that gives it away, it does this when its version is more correct than yours. If taste were the problem, a truer word would be a better outcome. When the text is fixed, it's still wrong.

There's no copy button

[0:49] Here's why these models have no copy button. When you ask for the whole paragraph back, every untouched word is repredicted one at a time from scratch. Each of those predictions usually lands on your original word. Usually is not always, and a paragraph is hundreds of chances.

Where the watermark comes in

[1:06] Watermarking can add a second push, nudging the model toward a marked subset of words at every step. One flag, whether a given watermark does that depends on its design, and some are built to leave the odds untouched. This is the seam where it depends is doing real work. Which is how let a team of six comes back as managed a team of six. Not a correction, not an improvement, just a word you didn't choose.

What it proves, and what it doesn't

[1:29] So text coming back unchanged doesn't prove it was protected. It may have been regenerated and simply landed on your words again. And a changed word doesn't prove the model judged yours worse. It may be drift or a nudge, neither of which read your writing at all. Which points at the fix. Let the model propose edits and let ordinary software apply them, so untouched text never enters the generator at all. The problem isn't that the model picked a worse word. It's that it was allowed to pick a word at all. Your turn. Here's the prompt. Read it

Your turn

[2:00] With me. Take something you wrote and care about. Ask for one specific fix and ask for the answer as a list of changes with locations, not as the written text. Then diff it against your original and count every word that moved. The count is almost never zero and the words that moved were never named in your request. Lay 'em in for bare. Why an AI can change a word you chose on purpose. Lay 'em in for bare.